Cascading Signal-Model Complexity for Energy-Aware Detection

Abstract

Equipping everyday objects with sensing and computational capabilities creates the potential to achieve context and situational awareness through the detection of natural events. A major challenge in energy-constrained devices is that the detection of natural events generally demands sophisticated signal modeling and processing, along with continuous sensing. In this paper, we propose to use a cascade architecture to control signal-model complexity; this approach allows the device to sense continuously and trigger a more accurate signal model only when an event of interest is likely to be occurring. The sensitivity of the triggering affects detection performance and device energy consumption, so we formulate and solve the problem of optimal threshold allocation to control this sensitivity. We prove that for controlling signal-model complexity, triggering as often as possible maximizes detection performance; this seemingly intuitive result does not hold in other popular cascading techniques, most notably in low-power wakeup mechanisms. Our analysis leads to a simple threshold-tracking algorithm that can adjust to time-varying environmental conditions and energy supply. Combining a low-power MCU with controllable supply-voltage scaling, we present an acoustic wildlife monitoring system that exhibits 12× energy scalability from cascading detectors, and an additional 2× from synergistic hardware scaling on commercially available components. In practical scenarios, we demonstrate energy savings ranging from 3× to 20× with minimal loss in performance, compared to the conventional approach.